Scheduling method and device for edge heterogeneous resources

Abstract

The invention discloses a resource scheduling method and device for edge devices, and the method comprises the steps: constructing an edge device performance prediction network library based on a random network generator and all edge devices, and when a target intelligent task is received, carrying out the prediction of the performance of the edge devices; determining a matched target edge device from all the edge devices based on an edge device performance prediction network library, scheduling the target intelligent task to the target edge device, and monitoring the device states of all the edge devices and the task operation states of all the target intelligent tasks operated by the edge devices, based on the monitoring results of all the edge devices, determining the running load capacity of each edge device, determining a low-load edge device based on the running load capacity, and scheduling a target intelligent task running on the low-load edge device to the determined corresponding to-be-migrated device; according to the method, the matching degree between the task and the edge heterogeneous resource is improved, so that the edge heterogeneous resource is fully utilized, and the efficiency of the edge heterogeneous resource is improved.

Description

Technical Field

[0001] The present invention relates to the field of edge computing technology, and in particular to a scheduling method and device for edge heterogeneous resources. Background Art

[0002] With the development of edge computing, more and more computing tasks are distributed to edge nodes close to data sources to reduce latency, reduce bandwidth consumption, and improve data processing efficiency. However, since edge nodes are usually heterogeneous (including differences in computing power, storage resources, network bandwidth, etc.) and dynamic (resource status changes at any time, and the network connection of edge nodes may be unstable), how to reasonably schedule tasks to improve the overall performance of the system has become a key challenge.

[0003] Existing edge computing scheduling algorithms are difficult to adapt to resource heterogeneity, dynamics, and performance requirements of different tasks at the same time. Especially in offline or weakly connected scenarios, existing methods often have problems with uneven resource allocation or load imbalance, resulting in low computing resource utilization or task execution failure. For example, scheduling methods based on heuristic algorithms such as shortest job first (SJF) are only suitable for simple scheduling environments. Their fixed strategies cannot adapt to the resource heterogeneity and dynamics of edge nodes, and it is difficult to make adaptive adjustments based on real-time resource status, resulting in low resource utilization or task execution failure. For another example, although the load balancing algorithm used for edge computing can disperse task pressure to a certain extent and avoid overloading of some nodes, the algorithm also performs poorly in heterogeneous device resource scheduling, and it is difficult to make accurate scheduling based on the resource characteristics of the node. For example, a task that requires a large amount of GPU resources is assigned to a node with a powerful CPU. Although the node has a low load, the execution efficiency may be very poor and heterogeneous resources cannot be fully utilized. In addition, existing scheduling algorithms based on machine learning mostly focus on the initial allocation of tasks, lack an adaptive adjustment mechanism for the dynamic changes of resource status over time, and are unable to cope with dynamic resource changes at edge nodes.

[0004] Therefore, it is particularly important to provide a scheduling solution for edge heterogeneous resources to improve the matching degree between tasks and edge heterogeneous resources, and then make full use of edge heterogeneous resources and improve the efficiency of edge heterogeneous resources. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a scheduling method and device for edge heterogeneous resources, which can improve the matching degree between tasks and edge heterogeneous resources, thereby making full use of edge heterogeneous resources and improving the efficiency of edge heterogeneous resources.

[0006] In order to solve the above technical problems, the first aspect of the present invention discloses a scheduling method for edge heterogeneous resources, the method comprising:

[0007] Building an edge device performance prediction network library based on a random network generator and all edge devices; the edge device performance prediction network library includes a performance prediction model corresponding to each edge device;

[0008] When a target intelligent task is received, a target edge device matching the target intelligent task is determined from all the edge devices based on the edge device performance prediction network library, and the target intelligent task is scheduled to the target edge device;

[0009] Monitor the device status of all the edge devices and the task running status of all the target intelligent tasks running on them, and obtain monitoring results of all the edge devices;

[0010] Based on the monitoring results of all the edge devices, determine the operating load of each edge device, and determine the low-load edge devices that have an operating load and an operating load lower than a preset operating load threshold from all the edge devices;

[0011] For each of the low-load edge devices, a device to be migrated corresponding to the low-load edge device is determined, and all the target intelligent tasks running on the low-load edge device are scheduled to the device to be migrated corresponding to the low-load edge device.

[0012] As an optional implementation, in the first aspect of the present invention, the step of building an edge device performance prediction network library based on a random network generator and all edge devices includes:

[0013] Generate a number of intelligent task network structures based on a random network generator, input all of the intelligent task network structures into each edge device, and obtain operating parameters of each of the intelligent task network structures running on each of the edge devices; the operating parameters at least include inference delay parameters, power consumption parameters, and memory occupancy parameters;

[0014] For each of the edge devices, an initial performance prediction model corresponding to the edge device is trained based on the operating parameters of all the intelligent task network structures running on the edge device to obtain a performance prediction model corresponding to the edge device;

[0015] Based on the performance prediction models corresponding to all the edge devices, an edge device performance prediction network library is constructed.

[0016] As an optional implementation, in the first aspect of the present invention, when a target intelligent task is received, a target edge device matching the target intelligent task is determined from all the edge devices based on the edge device performance prediction network library, including:

[0017] When a target intelligent task is received, all candidate edge devices matching the target intelligent task are determined from all the edge devices, and target performance prediction models corresponding to all the candidate edge devices are determined from the edge device performance prediction network library;

[0018] The target intelligent task is input into all the target performance prediction models for prediction, and the predicted operation parameters of the target intelligent task running on the selected edge device corresponding to each target performance prediction model are obtained; the predicted operation parameters at least include predicted inference delay parameters, predicted power consumption parameters and predicted memory occupancy parameters;

[0019] Based on at least one predicted running parameter of the target intelligent task running on each of the edge devices to be selected, all the edge devices to be selected are sorted to obtain a target edge device sequence;

[0020] Based on the target edge device sequence, a target edge device matching the target intelligent task is determined.

[0021] As an optional implementation, in the first aspect of the present invention, the sorting of all the edge devices to be selected based on at least one predicted operating parameter of the target intelligent task running on each of the edge devices to be selected includes:

[0022] For each of the edge devices to be selected, a weighted sum is performed based on each of the predicted operating parameters of the target intelligent task running on the edge device to be selected and a weight parameter corresponding to each of the predicted operating parameters to obtain a weighted evaluation parameter of the edge device to be selected; all the edge devices to be selected are sorted based on the weighted evaluation parameter of each edge device to be selected;

[0023] Alternatively, all the edge devices to be selected are sorted based on the predicted power consumption parameters of the target intelligent task running on each of the edge devices to be selected;

[0024] Alternatively, for each of the edge devices to be selected, based on each of the predicted operating parameters of the target intelligent task running on the edge device to be selected, the predicted margin resource parameter corresponding to the edge device to be selected is determined; based on the predicted margin resource parameter of each of the edge devices to be selected, all the edge devices to be selected are sorted.

[0025] As an optional implementation, in the first aspect of the present invention, for each of the low-load edge devices, determining a device to be migrated corresponding to the low-load edge device, and scheduling all the target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device, includes:

[0026] For each of the low-load edge devices, based on the operating load of the low-load edge device and the operating load of all other edge devices, determine all migratable devices corresponding to the low-load edge device from all other edge devices; determine the predicted operating parameters of all the target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device; based on the predicted operating parameters of all the target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device, determine the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device, and schedule all the target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device.

[0027] As an optional implementation, in the first aspect of the present invention, the predicted operation parameters of all the target intelligent tasks running on the low-load edge device running on each migratable device corresponding to the low-load edge device and the operation parameters of all the target intelligent tasks running on each migratable device corresponding to the low-load edge device, determining the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device, includes:

[0028] Determine the operating parameters of all the target intelligent tasks running on each migratable device corresponding to the low-load edge device; obtain comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device based on the predicted operating parameters of all the target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device and the operating parameters of all the target intelligent tasks running on each migratable device corresponding to the low-load edge device; determine the device to be migrated corresponding to the low-load edge device from all the migratable devices corresponding to the low-load edge device based on the comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device.

[0029] As an optional implementation, in the first aspect of the present invention, before determining, for each of the low-load edge devices, a device to be migrated corresponding to the low-load edge device, the method further includes:

[0030] All the low-load edge devices are sorted according to the operating load of each of the low-load edge devices to obtain a low-load edge device sequence; for all the low-load edge devices in the low-load edge device sequence, the operating load of the low-load edge device with the previous sequence number is lower than the operating load of the low-load edge device with the following sequence number;

[0031] And, for each of the low-load edge devices, based on the operating load of the low-load edge device and the operating load of all other edge devices, determining all migratable devices corresponding to the low-load edge device from all other edge devices, including:

[0032] For each of the low-load edge devices, according to the sequence number of the low-load edge device in the low-load edge device sequence, based on the operating load of the low-load edge device and the operating load of all other edge devices, all migratable devices corresponding to the low-load edge device are determined from all other edge devices.

[0033] A second aspect of the present invention discloses a scheduling device for edge heterogeneous resources, the device comprising:

[0034] A construction module, used to construct an edge device performance prediction network library based on a random network generator and all edge devices; the edge device performance prediction network library includes a performance prediction model corresponding to each edge device;

[0035] A first determination module is used for, when receiving a target intelligent task, determining a target edge device matching the target intelligent task from all the edge devices based on the edge device performance prediction network library, and scheduling the target intelligent task to the target edge device;

[0036] A monitoring module, used to monitor the device status of all the edge devices and the task running status of all the target intelligent tasks running on them, and obtain monitoring results of all the edge devices;

[0037] A second determination module is used to determine the operating load of each edge device based on the monitoring results of all the edge devices, and determine the low-load edge devices that have an operating load and an operating load lower than a preset operating load threshold from all the edge devices;

[0038] The migration module is used to determine, for each low-load edge device, a device to be migrated corresponding to the low-load edge device, and schedule all the target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device.

[0039] As an optional implementation, in the second aspect of the present invention, the construction module constructs an edge device performance prediction network library based on a random network generator and all edge devices, specifically in the following manner:

[0040] Generate a number of intelligent task network structures based on a random network generator, input all of the intelligent task network structures into each edge device, and obtain operating parameters of each of the intelligent task network structures running on each of the edge devices; the operating parameters at least include inference delay parameters, power consumption parameters, and memory occupancy parameters;

[0041] For each of the edge devices, an initial performance prediction model corresponding to the edge device is trained based on the operating parameters of all the intelligent task network structures running on the edge device to obtain a performance prediction model corresponding to the edge device;

[0042] Based on the performance prediction models corresponding to all the edge devices, an edge device performance prediction network library is constructed.

[0043] As an optional implementation, in the second aspect of the present invention, when the first determination module receives the target intelligent task, based on the edge device performance prediction network library, the target edge device matching the target intelligent task is determined from all the edge devices, specifically in the following manners:

[0044] When a target intelligent task is received, all candidate edge devices matching the target intelligent task are determined from all the edge devices, and target performance prediction models corresponding to all the candidate edge devices are determined from the edge device performance prediction network library;

[0045] The target intelligent task is input into all the target performance prediction models for prediction, and the predicted operation parameters of the target intelligent task running on the selected edge device corresponding to each target performance prediction model are obtained; the predicted operation parameters at least include predicted inference delay parameters, predicted power consumption parameters and predicted memory occupancy parameters;

[0046] Based on at least one predicted running parameter of the target intelligent task running on each of the edge devices to be selected, all the edge devices to be selected are sorted to obtain a target edge device sequence;

[0047] Based on the target edge device sequence, a target edge device matching the target intelligent task is determined.

[0048] As an optional implementation, in the second aspect of the present invention, the first determination module sorts all the edge devices to be selected based on at least one predicted operation parameter of the target intelligent task running on each of the edge devices to be selected, and the specific method includes:

[0049] For each of the edge devices to be selected, a weighted sum is performed based on each of the predicted operating parameters of the target intelligent task running on the edge device to be selected and a weight parameter corresponding to each of the predicted operating parameters to obtain a weighted evaluation parameter of the edge device to be selected; all the edge devices to be selected are sorted based on the weighted evaluation parameter of each edge device to be selected;

[0050] Alternatively, all the edge devices to be selected are sorted based on the predicted power consumption parameters of the target intelligent task running on each of the edge devices to be selected;

[0051] Alternatively, for each of the edge devices to be selected, based on each of the predicted operating parameters of the target intelligent task running on the edge device to be selected, the predicted margin resource parameter corresponding to the edge device to be selected is determined; based on the predicted margin resource parameter of each of the edge devices to be selected, all the edge devices to be selected are sorted.

[0052] As an optional implementation, in the second aspect of the present invention, the migration determination module determines, for each of the low-load edge devices, a device to be migrated corresponding to the low-load edge device, and schedules all the target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device, specifically in the following manners:

[0053] For each of the low-load edge devices, based on the operating load of the low-load edge device and the operating load of all other edge devices, determine all migratable devices corresponding to the low-load edge device from all other edge devices; determine the predicted operating parameters of all the target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device; based on the predicted operating parameters of all the target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device, determine the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device, and schedule all the target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device.

[0054] As an optional implementation, in the second aspect of the present invention, the migration module determines the device to be migrated corresponding to the low-load edge device from all the migratable devices corresponding to the low-load edge device based on the predicted operating parameters of all the target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device and the operating parameters of all the target intelligent tasks running on each migratable device corresponding to the low-load edge device, and the specific method includes:

[0055] Determine the operating parameters of all the target intelligent tasks running on each migratable device corresponding to the low-load edge device; obtain comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device based on the predicted operating parameters of all the target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device and the operating parameters of all the target intelligent tasks running on each migratable device corresponding to the low-load edge device; determine the device to be migrated corresponding to the low-load edge device from all the migratable devices corresponding to the low-load edge device based on the comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device.

[0056] As an optional implementation, in the second aspect of the present invention, the device further includes:

[0057] A sorting module, used for sorting all the low-load edge devices according to the operating load of each low-load edge device to obtain a low-load edge device sequence before the migration module determines the device to be migrated corresponding to each low-load edge device; for all the low-load edge devices in the low-load edge device sequence, the operating load of the low-load edge device with the previous sequence number is lower than the operating load of the low-load edge device with the following sequence number;

[0058] And, the migration module determines, for each of the low-load edge devices, all the migratable devices corresponding to the low-load edge device from all the other edge devices based on the operating load of the low-load edge device and the operating load of all the other edge devices, specifically in the following manners:

[0059] For each of the low-load edge devices, according to the sequence number of the low-load edge device in the low-load edge device sequence, based on the operating load of the low-load edge device and the operating load of all other edge devices, all migratable devices corresponding to the low-load edge device are determined from all other edge devices.

[0060] The third aspect of the present invention discloses another scheduling device for edge heterogeneous resources, the device comprising:

[0061] A memory storing executable program code;

[0062] a processor coupled to the memory;

[0063] The processor calls the executable program code stored in the memory to execute part or all of the steps in the scheduling method for edge heterogeneous resources disclosed in the first aspect of the present invention.

[0064] The fourth aspect of the present invention discloses a computer-storable medium, which stores computer instructions. When the computer instructions are called, they are used to execute some or all of the steps in the scheduling method for edge heterogeneous resources disclosed in the first aspect of the present invention.

[0065] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0066] The present invention can build an edge device performance prediction network library based on a random network generator and all edge devices. When a target intelligent task is received, a target edge device matching the target intelligent task is determined from all edge devices based on the edge device performance prediction network library, and the target intelligent task is scheduled to the target edge device. The device status of all edge devices and the task running status of all target intelligent tasks running on them are monitored to obtain the monitoring results of all edge devices. Based on the monitoring results of all edge devices, the running load of each edge device is determined, and low-load edge devices with running loads and running loads lower than a preset running load threshold are determined from all edge devices. For each low-load edge device, determine the device to be migrated corresponding to the low-load edge device, and schedule all target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device; it can be seen that the present invention can first predict the task based on the edge device performance prediction network library when allocating tasks, determine the target edge device according to the prediction result and schedule the task to the target edge device, and determine the low-load edge device according to the running load of each edge device and schedule the task on the low-load edge device to the corresponding device to be migrated, which is conducive to improving the matching degree between tasks and edge heterogeneous resources, and then making full use of edge heterogeneous resources and improving the efficiency of edge heterogeneous resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a flow chart of a scheduling method for edge heterogeneous resources disclosed in an embodiment of the present invention;

[0068] Figure 2 It is a schematic diagram of the process of constructing and extracting features of a random network generator disclosed in an embodiment of the present invention;

[0069] Figure 3 It is a schematic diagram of the prediction process of the edge device performance prediction network library disclosed in an embodiment of the present invention;

[0070] Figure 4 It is a flow chart of a two-stage scheduling based on prediction and adjustment for edge heterogeneous resources disclosed in an embodiment of the present invention;

[0071] Figure 5It is a timing relationship diagram of the execution of the dynamic adjustment scheduling stage disclosed in the embodiment of the present invention;

[0072] Figure 6 It is a structural schematic diagram of a scheduling device for edge heterogeneous resources disclosed in an embodiment of the present invention;

[0073] Figure 7 It is a structural schematic diagram of another scheduling device for edge heterogeneous resources disclosed in an embodiment of the present invention;

[0074] Figure 8 It is a structural schematic diagram of another scheduling device for edge heterogeneous resources disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0075] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0076] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, product or end including a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units inherent to these processes, methods, products or ends.

[0077] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0078] The present invention discloses a scheduling method and device for edge heterogeneous resources, which can predict tasks based on the edge device performance prediction network library when allocating tasks, determine the target edge device according to the prediction result and schedule the task to the target edge device, and determine the low-load edge device according to the running load of each edge device and schedule the task on the low-load edge device to the corresponding device to be migrated, which is conducive to improving the matching degree between tasks and edge heterogeneous resources, and then making full use of edge heterogeneous resources and improving the efficiency of edge heterogeneous resources. The following are detailed descriptions.

[0079] Embodiment 1

[0080] See also Figure 1 , Figure 1 : is a flow chart of a scheduling method for edge heterogeneous resources disclosed in an embodiment of the present invention. Figure 1 The method shown can be applied to scheduling scenarios for edge heterogeneous resources, and can also be applied to any resource scheduling scenario, which is not limited in the embodiments of the present invention. Figure 1 As shown, the scheduling method for edge heterogeneous resources may include the following operations:

[0081] 101. Building an edge device performance prediction network library based on a random network generator and all edge devices; the edge device performance prediction network library includes a performance prediction model corresponding to each edge device;

[0082] 102. When a target intelligent task is received, a target edge device matching the target intelligent task is determined from all edge devices based on the edge device performance prediction network library, and the target intelligent task is scheduled to the target edge device;

[0083] 103. Monitor the device status of all edge devices and the task running status of all target intelligent tasks running thereon, and obtain monitoring results of all edge devices;

[0084] In the embodiment of the present invention, it can be understood that since 101-102 selects the target edge device that matches the target intelligent task and schedules the target intelligent task to the target edge device, there is a situation where a single edge device runs a single target intelligent task, which will cause multiple edge devices to run at the same time with low load, resulting in equipment resource fragmentation and low equipment utilization. Through custom monitoring data items, edge device information can be collected at fixed time intervals to provide basic data support for subsequent dynamic scheduling.

[0085] 104. Based on the monitoring results of all edge devices, determine the operating load of each edge device, and determine the low-load edge devices that have operating loads and whose operating loads are lower than a preset operating load threshold from all edge devices;

[0086] 105. For each low-load edge device, determine the device to be migrated corresponding to the low-load edge device, and schedule all target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device.

[0087] It can be seen that the present invention can first predict the task based on the edge device performance prediction network library when allocating tasks, determine the target edge device according to the prediction results and schedule the task to the target edge device, and determine the low-load edge device according to the operating load of each edge device and schedule the task on the low-load edge device to the corresponding device to be migrated, which is conducive to improving the matching degree between tasks and edge heterogeneous resources, and then making full use of edge heterogeneous resources and improving the efficiency of edge heterogeneous resources.

[0088] In an optional embodiment, building an edge device performance prediction network library based on a random network generator and all edge devices may include:

[0089] Generate several intelligent task network structures based on a random network generator, input all intelligent task network structures into each edge device, and obtain the operating parameters of each intelligent task network structure running on each edge device; the operating parameters at least include inference delay parameters, power consumption parameters and memory usage parameters;

[0090] For each edge device, an initial performance prediction model corresponding to the edge device is trained based on the operating parameters of all intelligent task network structures running on the edge device to obtain a performance prediction model corresponding to the edge device;

[0091] Based on the performance prediction models corresponding to all edge devices, an edge device performance prediction network library is built.

[0092] In this optional embodiment, it is understandable that due to different hardware architectures and characteristics, the memory usage, inference delay, and power consumption of the same target intelligent task on different edge devices (edge ​​heterogeneous resources) are different. Random allocation may not meet the real-time requirements of intelligent tasks running on edge devices, and cannot fully utilize the characteristics of different edge devices. For example, a task that requires a large amount of GPU resources is assigned to a node with a powerful CPU. Although the node has a low load, the execution efficiency may be very poor. Therefore, an edge device performance prediction network library can be built to complete the resource requirements and performance status information of the target intelligent task running on different edge devices to guide the distribution of tasks.

[0093] In this optional embodiment, if Figure 2As shown in the figure, in the process of constructing the random network generator, common operator types include: convolution, full connection, pooling, etc., and common network structure types include single branch, multi-branch, residual network, etc. According to these characteristics, different operators are combined to construct single branch, multi-branch, and residual network random generators. Among them, the residual network needs to consider two types: BasicBlock and Bottleneck. By constructing different feature map size lists, channel number lists, and network layer operator type lists, they are combined to generate a calculation graph. Further optional, the characteristic parameters of a single layer of a neural network include convolution type, pooling type, full connection type, convolution step size, input channel number, output channel number, input neuron number, output neuron number, etc.; the characteristic parameters of the neural network structure include network depth, network layer type, and network layer connection method. Among them, the above-mentioned single-layer characteristic parameters can be combined to calculate the network's computational amount and parameter amount, which can reflect the network's time complexity and space complexity, thereby further completing the characterization of network inference latency and video memory occupancy. Each layer is represented by [computational amount, parameter amount] to retain the characteristics of the entire network layer, and each layer has multiple branch structures to ensure the unified representation of structures such as residual networks. At the same time, to facilitate the training of performance prediction models, the network feature list is uniformly filled with 0.

[0094] In this optional embodiment, the process of building an edge device performance prediction network library can be understood as extracting features from the above-mentioned random network, and selecting [number of cores, batchsize] allowed by the edge device parameters as input data, and inputting both information into the prediction network corresponding to a single edge device, so as to obtain information such as memory usage, inference latency, and power consumption of specific intelligent tasks under different parameters of different edge devices, thereby further optimizing the performance prediction model design. Among them, the performance prediction model includes a Transformer-based prediction model, and also a prediction model that uses a residual structure to extract network features and then connects to a multi-layer perceptron (MLP). The above-mentioned performance prediction models designed for different edge devices and the device numbers are stored together as key-values ​​to form an edge device performance prediction network library to provide support for subsequent prediction scheduling.

[0095] It can be seen that this optional embodiment can generate several intelligent task network structures based on a random network generator, input all intelligent task network structures into each edge device, obtain the operating parameters of each intelligent task network structure running on each edge device, and train based on the operating parameters of all intelligent task network structures running on the edge device to obtain the performance prediction model corresponding to the edge device; it is beneficial to improve the accuracy of the performance prediction model, and then improve the accuracy of building the edge device performance prediction network library, so as to improve the accuracy of the operating parameters of the intelligent task network structure running on each edge device, thereby improving the scheduling accuracy of the target intelligent task.

[0096] In another optional embodiment, when a target intelligent task is received, determining a target edge device matching the target intelligent task from all edge devices based on the edge device performance prediction network library may include:

[0097] When a target intelligent task is received, all candidate edge devices matching the target intelligent task are determined from all edge devices, and target performance prediction models corresponding to all candidate edge devices are determined from the edge device performance prediction network library;

[0098] Input the target intelligent task into all target performance prediction models for prediction, and obtain the predicted operation parameters of the target intelligent task running on the selected edge device corresponding to each target performance prediction model; the predicted operation parameters at least include predicted inference delay parameters, predicted power consumption parameters and predicted memory usage parameters;

[0099] Based on at least one predicted operating parameter of the target intelligent task running on each of the candidate edge devices, all candidate edge devices are sorted to obtain a target edge device sequence;

[0100] Based on the target edge device sequence, determine the target edge device that matches the target intelligent task.

[0101] In this optional embodiment, if Figure 3 As shown, the target intelligent task is input into all target performance prediction models for prediction, and the predicted operating parameters of the target intelligent task on the selected edge device corresponding to each target performance prediction model are obtained. This can be understood as feature extraction of the actual target intelligent task to be executed, selecting the target performance prediction model that matches the selected edge device from the edge device performance prediction network library, inputting the number of device cores to be executed, batchsize and the extracted task features into the target performance prediction model, and outputting the predicted inference delay parameters, predicted power consumption parameters and predicted memory usage parameters corresponding to the selected edge device and other information.

[0102] It can be seen that this optional embodiment can determine all candidate edge devices that match the target intelligent task from all edge devices when receiving the target intelligent task, and determine the target performance prediction models corresponding to all candidate edge devices from the edge device performance prediction network library; input the target intelligent task into all target performance prediction models for prediction, and obtain the predicted operating parameters of the target intelligent task running on the candidate edge device corresponding to each target performance prediction model; based on at least one predicted operating parameter of the target intelligent task running on each candidate edge device, all candidate edge devices are sorted to obtain a target edge device sequence; based on the target edge device sequence, the target edge device that matches the target intelligent task is determined, which is conducive to improving the determination accuracy of the predicted operating parameters, and then improving the determination accuracy of the target edge device sequence to improve the determination accuracy of the target edge device, thereby improving the scheduling accuracy of the target intelligent task.

[0103] In yet another optional embodiment, sorting all the candidate edge devices based on at least one predicted operating parameter of the target intelligent task running on each candidate edge device may include:

[0104] For each edge device to be selected, a weighted sum is performed based on each predicted operating parameter of the target intelligent task running on the edge device to be selected and a weight parameter corresponding to each predicted operating parameter to obtain a weighted evaluation parameter of the edge device to be selected; based on the weighted evaluation parameter of each edge device to be selected, all edge devices to be selected are sorted;

[0105] Alternatively, all the candidate edge devices are ranked based on predicted power consumption parameters of the target intelligent task running on each candidate edge device;

[0106] Alternatively, for each edge device to be selected, based on each predicted operating parameter of the target intelligent task running on the edge device to be selected, the predicted margin resource parameter corresponding to the edge device to be selected is determined; based on the predicted margin resource parameter of each edge device to be selected, all edge devices to be selected are sorted.

[0107] It can be seen that this optional embodiment can perform weighted summation for each edge device to be selected based on each predicted operating parameter of the target intelligent task running on the edge device to be selected and the weight parameter corresponding to each predicted operating parameter to obtain the weighted evaluation parameter of the edge device to be selected; sort all edge devices to be selected based on the weighted evaluation parameter of each edge device to be selected; or sort all edge devices to be selected based on the predicted power consumption parameter of the target intelligent task running on each edge device to be selected; or, for each edge device to be selected, determine the predicted margin resource parameter corresponding to the edge device to be selected based on each predicted operating parameter of the target intelligent task running on the edge device to be selected; sort all edge devices to be selected based on the predicted margin resource parameter of each edge device to be selected, which is conducive to improving the accuracy of sorting all edge devices to be selected, and then improving the accuracy of determining the target edge device, thereby improving the scheduling accuracy of the target intelligent task.

[0108] In yet another optional embodiment, for each low-load edge device, determining a device to be migrated corresponding to the low-load edge device, and scheduling all target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device may include:

[0109] For each low-load edge device, based on the operating load of the low-load edge device and the operating load of all other edge devices, determine all migratable devices corresponding to the low-load edge device from all other edge devices; determine the predicted operating parameters for each migratable device corresponding to the low-load edge device for all target intelligent tasks running on the low-load edge device; based on the predicted operating parameters for each migratable device corresponding to the low-load edge device for all target intelligent tasks running on the low-load edge device, determine the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device, and schedule all target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device.

[0110] It can be seen that this optional embodiment can, for each low-load edge device, determine all the migratable devices corresponding to the low-load edge device from all other edge devices based on the operating load of the low-load edge device and the operating load of all other edge devices; determine the predicted operating parameters for all target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device; based on the predicted operating parameters for all target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device, determine the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device, and schedule all target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device; this is conducive to improving the accuracy of determining the devices to be migrated, and thereby improving the scheduling accuracy of all target intelligent tasks running on the low-load edge devices.

[0111] In yet another optional embodiment, based on the predicted operating parameters of all target intelligent tasks running on the low-load edge device running on each migratable device corresponding to the low-load edge device and the operating parameters of all target intelligent tasks running on each migratable device corresponding to the low-load edge device, determining the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device may include:

[0112] Determine the operating parameters of all target intelligent tasks running on each migratable device corresponding to the low-load edge device; obtain comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device based on the predicted operating parameters of all target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device and the operating parameters of all target intelligent tasks running on each migratable device corresponding to the low-load edge device; determine the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device based on the comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device.

[0113] In this optional embodiment, it can be understood that the process of determining the comprehensive evaluation parameters can determine the weight parameters of the two and perform weighted summation based on the weight parameters based on the predicted operating parameters of all target intelligent tasks running on the low-load edge device and each migratable device corresponding to the low-load edge device, as well as the operating parameters of all target intelligent tasks running on each migratable device corresponding to the low-load edge device. It can also be other comprehensive determination methods, which are not limited to the embodiments of the present invention.

[0114] It can be seen that this optional embodiment is capable of determining the operating parameters of all target intelligent tasks running on each migratable device corresponding to the low-load edge device; obtaining comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device based on the predicted operating parameters of all target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device and the operating parameters of all target intelligent tasks running on each migratable device corresponding to the low-load edge device; determining the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device based on the comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device; this is conducive to improving the accuracy of determining the device to be migrated, and thereby improving the scheduling accuracy of all target intelligent tasks running on the low-load edge device.

[0115] In yet another optional embodiment, before determining, for each low-load edge device, a device to be migrated corresponding to the low-load edge device, the method may further include:

[0116] All low-load edge devices are sorted according to the operating load of each low-load edge device to obtain a low-load edge device sequence; for all low-load edge devices in the low-load edge device sequence, the operating load of the low-load edge device with the first sequence number is lower than the operating load of the low-load edge device with the last sequence number;

[0117] And, for each low-load edge device, based on the operating load of the low-load edge device and the operating load of all other edge devices, determining all migratable devices corresponding to the low-load edge device from all other edge devices may include:

[0118] For each low-load edge device, according to the sequence number of the low-load edge device in the low-load edge device sequence, based on the operating load of the low-load edge device and the operating load of all other edge devices, all migratable devices corresponding to the low-load edge device are determined from all other edge devices.

[0119] In this optional embodiment, it can be understood that by performing task migration according to the sequence number of the low-load edge device in the low-load edge device sequence, the target intelligent task on the low-load edge device with lower load can be transferred preferentially, which can not only reduce the migration cost, but also reduce the number of running edge devices and reduce power consumption. Furthermore, after all migratable devices corresponding to the low-load edge device are determined from all other edge devices based on the serial number of the low-load edge device in the low-load edge device sequence and the operating load of the low-load edge device and all other edge devices, the complete process of determining the device to be migrated and migrating the target intelligent task on the low-load edge device to the device to be migrated in the above-mentioned optional embodiment will continue to be executed, and the migration task of the next low-load edge device in the low-load edge device sequence will continue to be executed after the migration of the target intelligent task in the low-load edge device is completed; it can be further understood that after the migration task of the low-load edge device is completed, the low-load edge device will be deleted from the low-load edge device sequence, and if the device to be migrated corresponding to the low-load edge device is also in the low-load edge device sequence, when the load of the device to be migrated is higher than the operating load threshold after the migration task of the low-load edge device is completed, the device to be migrated will also be deleted from the low-load edge device sequence.

[0120] It can be seen that this optional embodiment can sort all low-load edge devices according to the operating load of each low-load edge device to obtain a low-load edge device sequence; for all low-load edge devices in the low-load edge device sequence, the operating load of the low-load edge device with the previous sequence number is lower than the operating load of the low-load edge device with the subsequent sequence number; for each low-load edge device, according to the sequence number of the low-load edge device in the low-load edge device sequence, based on the operating load of the low-load edge device and the operating load of all other edge devices, all migratable devices corresponding to the low-load edge device are determined from all other edge devices; this is conducive to improving the migration accuracy of task migration of low-load edge devices, which can not only reduce the cost of task migration, but also reduce the number of edge device operations.

[0121] In the above optional embodiments, it can be understood that if Figure 4As shown in the figure, the entire scheduling scheme can be understood as two stages: prediction and adjustment. In the first stage, the performance of the target intelligent task on the edge device is predicted to give a scheduling sequence for task distribution (prediction scheduling stage), and in the second stage, a new scheduling device sequence is given for the migratable task by monitoring real-time information (dynamic adjustment scheduling stage). Specifically, in the first stage, a random network generator is used to generate different intelligent task network structures, and the operating parameters on the edge device are obtained through actual measurement. The performance prediction model is trained for different edge devices so that it can more accurately predict the operating parameters (resource requirements and operating status information) of the target intelligent task. The edge device performance prediction network library corresponding to the edge device is constructed according to the obtained performance prediction model. When a specific target intelligent task is assigned, the predicted operating parameters of the target intelligent task are obtained based on the edge device performance prediction network library, and a suitable target edge device sequence is given to complete the prediction scheduling of the first stage. In the second stage, the load of the edge device is monitored in real time, and the migratable device is given for the edge device with a low load (such as less than 30% of the device utilization rate) and the device to be migrated is further determined to complete the dynamic adjustment scheduling of the second stage.

[0122] In the above optional embodiments, it can be understood that the deployment of this scheduling scheme is oriented to heterogeneous edge computing environments (such as satellites, drones, smart factories, etc.), and is composed of a central control node and multiple edge nodes. The central control node is deployed in the cloud or ground station, and the edge nodes are deployed on heterogeneous hardware devices (such as GPU, NPU, FPGA, etc.). Each production task is encapsulated as a containerized application so that the task can be migrated and assigned to different hardware platforms for operation, and unified management is achieved through the Kubernetes cluster and KubeEdge platform, and the two-stage scheduling algorithm of the present invention is combined to achieve efficient scheduling and execution of tasks. Further optional, a Kubernetes cluster is built in the ground station or the cloud, and the master node components of Kubernetes (such as API Server, Scheduler, Controller Manager, etc.) are installed. These components are used for global scheduling and management of tasks in the cluster. Among them, the prediction scheduling stage can be integrated in the kubernetes initial scheduler. When the target intelligent task enters the scheduler, the random network generator and the edge device performance prediction network library are first called to evaluate the predicted operating parameters of the task on each edge device, and the predicted operating parameters are stored as task metadata, which will determine the initial allocation of the task. Deploy KubeEdge edge node agent (EdgeCore) on each edge device to enable it to communicate with the central control node. Use device plugins to register heterogeneous devices (such as GPU, NPU, FPGA, etc.) on each edge device. For edge devices that need to be monitored, you can customize Exporter to complete the collection. Since the target intelligent task is bound to the edge device at the beginning of the allocation, there will be a phenomenon that the operating load of some edge devices is too low and there are many resource fragments. In the dynamic adjustment scheduling stage, a key-value pair [low-load edge id, device id to be migrated] is given to the task. When the edge device is disconnected from the cloud center, kubeedge can call the interface to restart the task. When the communication is normal, kubernetes guides the migration of the task. For example Figure 5 As shown, the figure describes the timing relationship of the dynamic adjustment scheduling stage execution in the solution of the present invention.

[0123] In the above optional embodiments, the above embodiments can be used for task scheduling on heterogeneous devices on orbiting satellites. With the development of artificial intelligence applications, orbiting satellites are equipped with high-performance heterogeneous resources such as CPU, GPU, NPU, etc. to support real-time data processing requirements in scenarios such as smart transportation and smart agriculture. The resources on the satellite are limited, and it is necessary to efficiently utilize existing equipment and meet performance requirements such as low power consumption and real-time performance. The scheme of the present invention can form a two-stage scheduling scheme by predicting the demand for heterogeneous resources for specific tasks, which can effectively improve the rationality of task allocation and improve the utilization rate of limited resources. In addition, the above embodiments can also be used for quality inspection systems in smart factories. In the smart factory scenario, the quality inspection tasks in the factory may involve a large amount of image processing and AI reasoning, and the automation equipment on the production line (such as industrial robots, sensor processing units, etc.) is equipped with heterogeneous hardware resources. This solution is deployed in the production environment of the factory, and a custom scheduler is used to predict the processing capacity of each device, allocate production tasks to ensure that the tasks can run on the most suitable device, and dynamically adjust the allocation and execution of tasks according to the load of the equipment to reduce production delays and improve production efficiency.

[0124] Embodiment 2

[0125] See also Figure 6 , Figure 6 : is a schematic diagram of a scheduling device for edge heterogeneous resources disclosed in an embodiment of the present invention. Figure 6 The device shown can be applied to scheduling scenarios for edge heterogeneous resources, and can also be applied to any resource scheduling scenario, which is not limited in the embodiments of the present invention. Figure 6 As shown, the scheduling device for edge heterogeneous resources may include:

[0126] A construction module 201 is used to construct an edge device performance prediction network library based on a random network generator and all edge devices; the edge device performance prediction network library includes a performance prediction model corresponding to each edge device;

[0127] A first determination module 202 is used to, when receiving a target intelligent task, determine a target edge device matching the target intelligent task from all edge devices based on an edge device performance prediction network library, and schedule the target intelligent task to the target edge device;

[0128] The monitoring module 203 is used to monitor the device status of all edge devices and the task running status of all target intelligent tasks running on them, and obtain the monitoring results of all edge devices;

[0129] A second determination module 204 is used to determine the operating load of each edge device based on the monitoring results of all edge devices, and determine the low-load edge devices that have an operating load and an operating load lower than a preset operating load threshold from all edge devices;

[0130] The migration module 205 is used to determine, for each low-load edge device, a device to be migrated corresponding to the low-load edge device, and schedule all target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device.

[0131] It can be seen that the device described in this embodiment can first predict the task based on the edge device performance prediction network library when allocating tasks, determine the target edge device according to the prediction results and schedule the task to the target edge device, and determine the low-load edge device according to the operating load of each edge device and schedule the task on the low-load edge device to the corresponding device to be migrated, which is conducive to improving the matching degree between tasks and edge heterogeneous resources, and then making full use of edge heterogeneous resources and improving the efficiency of edge heterogeneous resources.

[0132] In an optional embodiment, the construction module 201 constructs an edge device performance prediction network library based on the random network generator and all edge devices, specifically in the following manners:

[0133] Generate several intelligent task network structures based on a random network generator, input all intelligent task network structures into each edge device, and obtain the operating parameters of each intelligent task network structure running on each edge device; the operating parameters at least include inference delay parameters, power consumption parameters and memory usage parameters;

[0134] For each edge device, an initial performance prediction model corresponding to the edge device is trained based on the operating parameters of all intelligent task network structures running on the edge device to obtain a performance prediction model corresponding to the edge device;

[0135] Based on the performance prediction models corresponding to all edge devices, an edge device performance prediction network library is built.

[0136] It can be seen that the device described in this embodiment can generate several intelligent task network structures based on a random network generator, input all intelligent task network structures into each edge device, obtain the operating parameters of each intelligent task network structure running on each edge device, and train based on the operating parameters of all intelligent task network structures running on the edge device to obtain the performance prediction model corresponding to the edge device; it is beneficial to improve the accuracy of the performance prediction model, and then improve the accuracy of building the edge device performance prediction network library, so as to improve the accuracy of the operating parameters of the intelligent task network structure running on each edge device, thereby improving the scheduling accuracy of the target intelligent task.

[0137] In another optional embodiment, when receiving the target intelligent task, the first determination module 202 determines the target edge device matching the target intelligent task from all edge devices based on the edge device performance prediction network library, specifically in the following manners:

[0138] When a target intelligent task is received, all candidate edge devices matching the target intelligent task are determined from all edge devices, and target performance prediction models corresponding to all candidate edge devices are determined from the edge device performance prediction network library;

[0139] Input the target intelligent task into all target performance prediction models for prediction, and obtain the predicted operation parameters of the target intelligent task running on the selected edge device corresponding to each target performance prediction model; the predicted operation parameters at least include predicted inference delay parameters, predicted power consumption parameters and predicted memory usage parameters;

[0140] Based on at least one predicted operating parameter of the target intelligent task running on each of the candidate edge devices, all candidate edge devices are sorted to obtain a target edge device sequence;

[0141] Based on the target edge device sequence, determine the target edge device that matches the target intelligent task.

[0142] It can be seen that the device described in the implementation of this embodiment can, when receiving the target intelligent task, determine all candidate edge devices that match the target intelligent task from all edge devices, and determine the target performance prediction models corresponding to all candidate edge devices from the edge device performance prediction network library; input the target intelligent task into all target performance prediction models for prediction, and obtain the predicted operating parameters of the target intelligent task running on the candidate edge device corresponding to each target performance prediction model; based on at least one predicted operating parameter of the target intelligent task running on each candidate edge device, all candidate edge devices are sorted to obtain a target edge device sequence; based on the target edge device sequence, the target edge device that matches the target intelligent task is determined, which is conducive to improving the determination accuracy of the predicted operating parameters, and then improving the determination accuracy of the target edge device sequence, so as to improve the determination accuracy of the target edge device, thereby improving the scheduling accuracy of the target intelligent task.

[0143] In another optional embodiment, the first determination module 202 sorts all the candidate edge devices based on at least one predicted operation parameter of the target intelligent task running on each candidate edge device, specifically in the following manner:

[0144] For each edge device to be selected, a weighted sum is performed based on each predicted operating parameter of the target intelligent task running on the edge device to be selected and a weight parameter corresponding to each predicted operating parameter to obtain a weighted evaluation parameter of the edge device to be selected; based on the weighted evaluation parameter of each edge device to be selected, all edge devices to be selected are sorted;

[0145] Alternatively, all the candidate edge devices are ranked based on predicted power consumption parameters of the target intelligent task running on each candidate edge device;

[0146] Alternatively, for each edge device to be selected, based on each predicted operating parameter of the target intelligent task running on the edge device to be selected, the predicted margin resource parameter corresponding to the edge device to be selected is determined; based on the predicted margin resource parameter of each edge device to be selected, all edge devices to be selected are sorted.

[0147] It can be seen that the device described in the embodiment performs weighted summation for each edge device to be selected based on each predicted operating parameter of the target intelligent task running on the edge device to be selected and the weight parameter corresponding to each predicted operating parameter to obtain the weighted evaluation parameter of the edge device to be selected; sorts all edge devices to be selected based on the weighted evaluation parameter of each edge device to be selected; or sorts all edge devices to be selected based on the predicted power consumption parameter of the target intelligent task running on each edge device to be selected; or, for each edge device to be selected, determines the predicted margin resource parameter corresponding to the edge device to be selected based on each predicted operating parameter of the target intelligent task running on the edge device to be selected; sorts all edge devices to be selected based on the predicted margin resource parameter of each edge device to be selected, which is conducive to improving the accuracy of sorting all edge devices to be selected, thereby improving the accuracy of determining the target edge device, thereby improving the scheduling accuracy of the target intelligent task.

[0148] In another optional embodiment, the migration module 205 determines, for each low-load edge device, a device to be migrated corresponding to the low-load edge device, and schedules all target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device, specifically in the following manners:

[0149] For each low-load edge device, based on the operating load of the low-load edge device and the operating load of all other edge devices, determine all migratable devices corresponding to the low-load edge device from all other edge devices; determine the predicted operating parameters for each migratable device corresponding to the low-load edge device for all target intelligent tasks running on the low-load edge device; based on the predicted operating parameters for each migratable device corresponding to the low-load edge device for all target intelligent tasks running on the low-load edge device, determine the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device, and schedule all target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device.

[0150] It can be seen that the device described in the implementation of this embodiment can, for each low-load edge device, determine all the migratable devices corresponding to the low-load edge device from all other edge devices based on the operating load of the low-load edge device and the operating load of all other edge devices; determine the predicted operating parameters of all target intelligent tasks running on the low-load edge device for each migratable device corresponding to the low-load edge device; determine the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device based on the predicted operating parameters of all target intelligent tasks running on the low-load edge device for each migratable device corresponding to the low-load edge device, and schedule all target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device; this is conducive to improving the accuracy of determining the devices to be migrated, and thereby improving the scheduling accuracy of all target intelligent tasks running on the low-load edge devices.

[0151] In another optional embodiment, the migration module 205 determines the device to be migrated corresponding to the low-load edge device from all the migratable devices corresponding to the low-load edge device based on the predicted operating parameters of all target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device and the operating parameters of all target intelligent tasks running on each migratable device corresponding to the low-load edge device, and the specific method includes:

[0152] Determine the operating parameters of all target intelligent tasks running on each migratable device corresponding to the low-load edge device; obtain comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device based on the predicted operating parameters of all target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device and the operating parameters of all target intelligent tasks running on each migratable device corresponding to the low-load edge device; determine the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device based on the comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device.

[0153] It can be seen that the device described in the implementation of this embodiment is capable of determining the operating parameters of all target intelligent tasks running on each migratable device corresponding to the low-load edge device; obtaining comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device based on the predicted operating parameters of all target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device and the operating parameters of all target intelligent tasks running on each migratable device corresponding to the low-load edge device; determining the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device based on the comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device; this is conducive to improving the accuracy of determining the device to be migrated, and thereby improving the scheduling accuracy of all target intelligent tasks running on the low-load edge device.

[0154] In yet another optional embodiment, Figure 7 As shown, the device may also include:

[0155] The sorting module 206 is used to sort all low-load edge devices according to the operating load of each low-load edge device to obtain a low-load edge device sequence before the migration module 205 determines the device to be migrated corresponding to the low-load edge device for each low-load edge device; for all low-load edge devices in the low-load edge device sequence, the operating load of the low-load edge device with the previous sequence number is lower than the operating load of the low-load edge device with the following sequence number;

[0156] And, for each low-load edge device, the migration module 205 determines all the migratable devices corresponding to the low-load edge device from all other edge devices based on the operating load of the low-load edge device and the operating load of all other edge devices, specifically in the following manners:

[0157] For each low-load edge device, according to the sequence number of the low-load edge device in the low-load edge device sequence, based on the operating load of the low-load edge device and the operating load of all other edge devices, all migratable devices corresponding to the low-load edge device are determined from all other edge devices.

[0158] It can be seen that the device described in the implementation of this embodiment can sort all low-load edge devices according to the operating load of each low-load edge device to obtain a low-load edge device sequence; for all low-load edge devices in the low-load edge device sequence, the operating load of the low-load edge device with the previous sequence number is lower than the operating load of the low-load edge device with the subsequent sequence number; for each low-load edge device, according to the sequence number of the low-load edge device in the low-load edge device sequence, based on the operating load of the low-load edge device and the operating load of all other edge devices, all migratable devices corresponding to the low-load edge device are determined from all other edge devices; this is conducive to improving the migration accuracy of the low-load edge device for task migration, which can not only reduce the cost of task migration, but also reduce the number of edge device operations.

[0159] Embodiment 3

[0160] See also Figure 8 , Figure 8 is a schematic diagram of the structure of another scheduling device for edge heterogeneous resources disclosed in an embodiment of the present invention. Figure 8 The described device can be applied in an application server. Figure 8 As shown, the device may include:

[0161] A memory 301 storing executable program codes;

[0162] a processor 302 coupled to the memory 301;

[0163] The processor 302 calls the executable program code stored in the memory 301 to execute the steps in the scheduling method for edge heterogeneous resources described in the first embodiment of the present invention.

[0164] Embodiment 4

[0165] An embodiment of the present invention discloses a computer storable medium, which stores computer instructions. When the computer instructions are called, they are used to execute the steps of the scheduling method for edge heterogeneous resources described in the first embodiment of the present invention.

[0166] Embodiment 5

[0167] An embodiment of the present invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to enable a computer to execute the steps in the scheduling method for edge heterogeneous resources described in the first embodiment of the present invention.

[0168] The device embodiments described above are only illustrative, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, i.e., they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0169] Through the specific description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution can be essentially or partly contributed to the prior art in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, a magnetic disk storage, a magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0170] Finally, it should be noted that the scheduling method and device for edge heterogeneous resources disclosed in the embodiments of the present invention only disclose the preferred embodiments of the present invention, which are only used to illustrate the technical solution of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A resource scheduling method for edge devices, characterized in that: The method comprises: Building an edge device performance prediction network library based on a random network generator and all edge devices; the edge device performance prediction network library includes a performance prediction model corresponding to each edge device; When a target intelligent task is received, a target edge device matching the target intelligent task is determined from all the edge devices based on the edge device performance prediction network library, and the target intelligent task is scheduled to the target edge device; Monitor the device status of all the edge devices and the task running status of all the target intelligent tasks running on them, and obtain monitoring results of all the edge devices; Based on the monitoring results of all the edge devices, determine the operating load of each edge device, and determine the low-load edge devices that have an operating load and an operating load lower than a preset operating load threshold from all the edge devices; For each of the low-load edge devices, a device to be migrated corresponding to the low-load edge device is determined, and all the target intelligent tasks running on the low-load edge device are scheduled to the device to be migrated corresponding to the low-load edge device.

2. The resource scheduling method for edge devices according to claim 1, characterized in that: The edge device performance prediction network library is constructed based on the random network generator and all edge devices, including: Generate a number of intelligent task network structures based on a random network generator, input all of the intelligent task network structures into each edge device, and obtain operating parameters of each of the intelligent task network structures running on each of the edge devices; the operating parameters at least include inference delay parameters, power consumption parameters, and memory occupancy parameters; For each of the edge devices, an initial performance prediction model corresponding to the edge device is trained based on the operating parameters of all the intelligent task network structures running on the edge device to obtain a performance prediction model corresponding to the edge device; Based on the performance prediction models corresponding to all the edge devices, an edge device performance prediction network library is constructed.

3. The resource scheduling method for edge devices according to claim 2, characterized in that: When receiving the target intelligent task, based on the edge device performance prediction network library, determining a target edge device matching the target intelligent task from all the edge devices includes: When a target intelligent task is received, all candidate edge devices matching the target intelligent task are determined from all the edge devices, and target performance prediction models corresponding to all the candidate edge devices are determined from the edge device performance prediction network library; The target intelligent task is input into all the target performance prediction models for prediction, and the predicted operation parameters of the target intelligent task running on the selected edge device corresponding to each target performance prediction model are obtained; the predicted operation parameters at least include predicted inference delay parameters, predicted power consumption parameters and predicted memory occupancy parameters; Based on at least one predicted running parameter of the target intelligent task running on each of the edge devices to be selected, all the edge devices to be selected are sorted to obtain a target edge device sequence; Based on the target edge device sequence, a target edge device matching the target intelligent task is determined.

4. The resource scheduling method for edge devices according to claim 3, characterized in that: The sorting of all the edge devices to be selected based on at least one predicted operating parameter of the target intelligent task running on each of the edge devices to be selected includes: For each of the edge devices to be selected, a weighted sum is performed based on each of the predicted operating parameters of the target intelligent task running on the edge device to be selected and a weight parameter corresponding to each of the predicted operating parameters to obtain a weighted evaluation parameter of the edge device to be selected; all the edge devices to be selected are sorted based on the weighted evaluation parameter of each edge device to be selected; Alternatively, all the edge devices to be selected are sorted based on the predicted power consumption parameters of the target intelligent task running on each of the edge devices to be selected; Alternatively, for each of the edge devices to be selected, based on each of the predicted operating parameters of the target intelligent task running on the edge device to be selected, the predicted margin resource parameter corresponding to the edge device to be selected is determined; based on the predicted margin resource parameter of each of the edge devices to be selected, all the edge devices to be selected are sorted.

5. The resource scheduling method for edge devices according to claim 4, characterized in that: The step of determining, for each of the low-load edge devices, a device to be migrated corresponding to the low-load edge device, and scheduling all the target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device comprises: For each of the low-load edge devices, based on the operating load of the low-load edge device and the operating load of all other edge devices, determine all migratable devices corresponding to the low-load edge device from all other edge devices; determine the predicted operating parameters of all the target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device; based on the predicted operating parameters of all the target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device, determine the device to be migrated corresponding to the low-load edge device from all migratable devices corresponding to the low-load edge device, and schedule all the target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device.

6. The resource scheduling method for edge devices according to claim 5, characterized in that: The method of determining the device to be migrated corresponding to the low-load edge device from all the migratable devices corresponding to the low-load edge device based on the predicted operating parameters of all the target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device and the operating parameters of all the target intelligent tasks running on each migratable device corresponding to the low-load edge device, comprises: Determine the operating parameters of all the target intelligent tasks running on each migratable device corresponding to the low-load edge device; obtain comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device based on the predicted operating parameters of all the target intelligent tasks running on the low-load edge device on each migratable device corresponding to the low-load edge device and the operating parameters of all the target intelligent tasks running on each migratable device corresponding to the low-load edge device; determine the device to be migrated corresponding to the low-load edge device from all the migratable devices corresponding to the low-load edge device based on the comprehensive evaluation parameters of each migratable device corresponding to the low-load edge device.

7. The resource scheduling method for edge devices according to claim 5, characterized in that: Before determining, for each of the low-load edge devices, a device to be migrated corresponding to the low-load edge device, the method further includes: All the low-load edge devices are sorted according to the operating load of each of the low-load edge devices to obtain a low-load edge device sequence; for all the low-load edge devices in the low-load edge device sequence, the operating load of the low-load edge device with the previous sequence number is lower than the operating load of the low-load edge device with the following sequence number; And, for each of the low-load edge devices, based on the operating load of the low-load edge device and the operating load of all other edge devices, determining all migratable devices corresponding to the low-load edge device from all other edge devices, including: For each of the low-load edge devices, according to the sequence number of the low-load edge device in the low-load edge device sequence, based on the operating load of the low-load edge device and the operating load of all other edge devices, all migratable devices corresponding to the low-load edge device are determined from all other edge devices.

8. A resource scheduling device for edge devices, characterized in that: The method comprises: A construction module, used to construct an edge device performance prediction network library based on a random network generator and all edge devices; the edge device performance prediction network library includes a performance prediction model corresponding to each edge device; A first determination module is used for, when receiving a target intelligent task, determining a target edge device matching the target intelligent task from all the edge devices based on the edge device performance prediction network library, and scheduling the target intelligent task to the target edge device; A monitoring module, used to monitor the device status of all the edge devices and the task running status of all the target intelligent tasks running on them, and obtain monitoring results of all the edge devices; A second determination module is used to determine the operating load of each edge device based on the monitoring results of all the edge devices, and determine the low-load edge devices that have an operating load and an operating load lower than a preset operating load threshold from all the edge devices; The migration module is used to determine, for each low-load edge device, a device to be migrated corresponding to the low-load edge device, and schedule all the target intelligent tasks running on the low-load edge device to the device to be migrated corresponding to the low-load edge device.

9. A resource scheduling device for edge devices, characterized in that: The device comprises: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the resource scheduling method for edge devices as described in any one of claims 1-7.

10. A computer storable medium, characterized in that: The computer storage medium stores computer instructions, and when the computer instructions are called, they are used to execute the resource scheduling method for edge devices as described in any one of claims 1-7.

Images